AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs

Ngoc Bui, Shubham Sharma, Simran Lamba, Saumitra Mishra, Rex Ying

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 16 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add forward KL distillation loss (ablation)TRIM-KV · AIME24 (budget 4096)pass@1: 72.1 -> 75.8Improved
add next-token prediction loss (ablation)TRIM-KV · AIME24 (budget 4096)pass@1: 72.5 -> 75.8Improved
add memory capacity loss (ablation)TRIM-KV · AIME24 (budget 4096)pass@1: 42.9 -> 75.8Improved
evict with TRIM-KV instead of SnapKV at 32768 budgetQwen3-4B-Instruct · LongMemEval_Saccuracy (%): 27.6 -> 44.8Improved
evict with TRIM-KV instead of SnapKV at 8192 budgetQwen3-4B-Instruct · LongMemEval_Saccuracy (%): 15.8 -> 34.0Improved
evict with TRIM-KV instead of LocRetPhi3-mini-128K · LongBench-V2average accuracy (%): 28.03 -> 34.09Improved
evict with TRIM-KV instead of LocRetPhi3-mini-128K · LongBenchaverage relative change (%): -4.82 -> -2.54Improved
use TRIM-KV instead of R-KV at 2048 budgetQwen3-4B · LongProc Pseudo-to-Code (2k)accuracy (%): 2.5 -> 18.0Improved
use TRIM-KV eviction instead of full cacheQwen3-4B · LongProc Countdown (2k)accuracy (%): 90.0 -> 93.5Improved
use TRIM-KV eviction instead of full cacheQwen3-4B · AIME24pass@1: 65.5 -> 75.8Improved
use TRIM-KV eviction instead of full cacheQwen3-4B · LongProc Pseudo-to-Code (2k)accuracy (%): 25.0 -> 18.0Hurt
apply TRIM-KV eviction on incompressible retrievalQwen3-4B-Instruct · SCBench Retr.KVscore: 66.00 -> 0.00Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy